[Kernel] Add MXFP4 W4A4 CUTLASS MoE kernel for SM100 (#37463)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-04-17 16:42:32 -07:00
committed by GitHub
parent 5cdddddd4a
commit a8bffaa133
11 changed files with 1701 additions and 14 deletions
+22
View File
@@ -116,6 +116,12 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
" Tensor a_blockscale, Tensor b_blockscales, Tensor alphas,"
" Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()");
// cutlass mxfp4 block scaled group GEMM (MXFP4 x MXFP4 MoE)
ops.def(
"cutlass_mxfp4_group_mm(Tensor! out, Tensor a, Tensor b,"
" Tensor a_blockscale, Tensor b_blockscales,"
" Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()");
// Compute NVFP4 block quantized tensor.
ops.def(
"scaled_fp4_quant(Tensor input,"
@@ -149,6 +155,19 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
"Tensor input, Tensor input_global_scale, Tensor input_offset_by_experts,"
"Tensor output_scale_offset_by_experts) -> ()");
// Compute MXFP4 experts quantization (32-element blocks, E8M0 SFs).
ops.def(
"mxfp4_experts_quant(Tensor! output, Tensor! output_scale,"
"Tensor input, Tensor input_offset_by_experts,"
"Tensor output_scale_offset_by_experts, int n_experts) -> ()");
// Fused SiLU+Mul+MXFP4 experts quantization.
ops.def(
"silu_and_mul_mxfp4_experts_quant(Tensor! output, Tensor! "
"output_scale,"
"Tensor input, Tensor input_offset_by_experts,"
"Tensor output_scale_offset_by_experts, int n_experts) -> ()");
// Fused SiLU+Mul+NVFP4 quantization.
ops.def(
"silu_and_mul_nvfp4_quant(Tensor! result, Tensor! result_block_scale, "
@@ -233,6 +252,9 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
ops.impl("silu_and_mul_scaled_fp4_experts_quant",
TORCH_BOX(&silu_and_mul_scaled_fp4_experts_quant));
ops.impl("silu_and_mul_nvfp4_quant", TORCH_BOX(&silu_and_mul_nvfp4_quant));
ops.impl("mxfp4_experts_quant", TORCH_BOX(&mxfp4_experts_quant));
ops.impl("silu_and_mul_mxfp4_experts_quant",
TORCH_BOX(&silu_and_mul_mxfp4_experts_quant));
// W4A8 ops: impl registrations are in the source files
// (w4a8_mm_entry.cu and w4a8_grouped_mm_entry.cu)